The goal of this project is to implement a quality-enhancing tool of a image using machine learning.
We use a Gaussian Mixture Model to learn a conversion fonction between a spline interpolated image and the desired image via EM. The training data is the concatenation of the rgb components of a pixel and it's 8-pixel neighborhood of an interpolated image of the same size. Training data is built from different layers of a gaussian pyramid and an bicubic-interpolation pyramid.
We then use the learned joint distribution to predict output pixels from it's neighborhood in a scaled interpolated image.
It implements a model of prediction proposed here Single Image Super-Resolution - He et al.
- The output is not satisfying, Some quality has been gained but it is roughly equivalent to aplying unsharp masking.
- The initialization of the gmm does not permit the EM to make it converge to an appropriate representqtion of the joint distribution. --> We obtain worse results than spline interpolation at PSNR and SSIM metrics